I've switched from sequential initialization of prepared statements on the connections to parallel initialization. That reduced startup time. I've also added code to reuse ForkJoinPool and keep it in memory for up tp 60s after the last prepared statement is initialized. That should allow it to be reused even if at some point we would decide to add new data sources at runtime (and it reduces load time as we do not need to created and destroy threads).

Time slashed by half:

pool=200 parallel: 141s serial: 338s

pool=default (32) parallel: 42s serial: 72s